VoxCPM

VoxCPM2 is an advanced open-source text-to-speech model developed by OpenBMB (Open Lab for Big Model Base). It is a tokenizer-free TTS system capable of multilingual speech generation, creative voice design, and highly realistic voice cloning. With over 33,200 GitHub stars and 3,800 forks, it is one of the most popular open-source speech synthesis projects.

The project is hosted at github.com/OpenBMB/VoxCPM. OpenBMB is a research lab affiliated with Tsinghua University and the Beijing Academy of Artificial Intelligence (BAAI), focused on building large-scale foundational AI models.

• Tokenizer-Free Architecture - Directly models raw audio without phonetic tokenizers, producing more natural prosody.
• Multilingual Generation - Supports multiple languages including English, Chinese, Japanese, Korean, and more.
• True-to-Life Cloning - Voice cloning with minimal reference audio that preserves natural intonation and emotion.
• Creative Voice Design - Generate novel voices with specified characteristics (pitch, timbre, speaking style).
• Zero-Shot Capability - Can generate speech for unseen speakers without fine-tuning.
• High-Fidelity Output - Produces studio-quality speech with natural pauses, emphasis, and emotional variation.

VoxCPM2 uses a tokenizer-free architecture that directly processes raw audio waveforms, avoiding the information loss and unnatural artifacts introduced by discrete tokenization methods common in traditional TTS systems. This approach enables the model to capture subtle acoustic details, prosodic nuances, and emotional inflections that tokenizer-based models typically miss.

The model is built on the CPM (Continual Pre-training Model) architecture, OpenBMB's family of large language models adapted for generative tasks. VoxCPM2 extends this architecture with specialized audio processing modules for generation, cloning, and voice design tasks.

VoxCPM2 excels at multilingual speech generation, supporting a wide range of languages with natural pronunciation and intonation. The model handles code-switching (mixing languages within a single utterance) without degradation, making it particularly useful for bilingual or multilingual content creation.

Supported languages include Mandarin Chinese, English, Japanese, Korean, Cantonese, and other major world languages. The tokenizer-free approach is especially beneficial for non-English languages, where tokenizer-based systems often struggle with phonetic accuracy.

VoxCPM2's voice cloning capability is one of its standout features. With as little as 3-10 seconds of reference audio, the model can capture a speaker's voice characteristics and generate new speech in that voice. The cloning preserves fine-grained details like accent, speaking rhythm, and emotional tone.

The Creative Voice Design feature takes this further, allowing users to blend voice characteristics, adjust pitch and timbre parameters, and create entirely new synthetic voices that don't correspond to any real person — useful for character voices, audiobooks, and creative content production.

git clone https://github.com/OpenBMB/VoxCPM.git
cd VoxCPM
pip install -r requirements.txt
# Download pretrained model weights
python scripts/download_model.py

Pre-trained model weights are available on Hugging Face. VoxCPM2 requires a CUDA-compatible GPU for inference. The examples/ directory contains usage examples for generation, cloning, and voice design workflows.